[flang]When flang-new compiles for the RISC-V platform, the object files have an incorrect ABI.

[buggfg]

I am cross-compiling the SPEC CPU2006 benchmarks using the LLVM 19.1.0 compiler on an Intel machine. The C/C++ examples can be successfully built into executable files using runspec and run successfully on the** RISC-V** platform. However, for Fortran examples like 410.bwaves, building with flang-new fails, and the error is:

ld.lld: error: bwaves.lto.o: cannot link object files with different floating-point ABI from ~/gnu-riscv64-20241106/sysroot/usr/lib/crt1.o.

PS. config info
**FOPTIMIZE **= -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math
**EXTRA_LDFLAGS **= -fuse-ld==ld.lld -L${gnu-riscv64}/lib/

Through debugging, I found that the ABI of bwaves.lto.o is rv64i2p1, while the ABI of the library file crt1.o is rv64i2p1_m2p0_a2p1_f2p2_d2p2_c2p0_zicsr2p0_zmmul1p0. Clearly, bwaves.lto.o is generated incorrectly. Could you please advise me on how to resolve this issue? Thank you very much!

The complete run script is：
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o block_solver.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 block_solver.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o flow_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 flow_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o flux_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 flux_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o jacobian_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 jacobian_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o shell_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 shell_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -fuse-ld=ld.lld -L{RISCV}/lib/ block_solver.o flow_lam.o flux_lam.o jacobian_lam.o shell_lam.o -o bwaves

[llvmbot]

@llvm/issue-subscribers-backend-risc-v

Author: bernadate (buggfg)

I am cross-compiling the **SPEC CPU2006** benchmarks using the **LLVM 19.1.0** compiler on an Intel machine. The C/C++ examples can be successfully built into executable files using runspec and run successfully on the** RISC-V** platform. However, for Fortran examples like **410.bwaves**, building with flang-new fails, and the error is:

ld.lld: error: bwaves.lto.o: cannot link object files with different floating-point ABI from ~/gnu-riscv64-20241106/sysroot/usr/lib/crt1.o.

PS. config info
**FOPTIMIZE **= -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math
**EXTRA_LDFLAGS **= -fuse-ld==ld.lld -L${gnu-riscv64}/lib/

Through debugging, I found that the ABI of bwaves.lto.o is rv64i2p1, while the ABI of the library file crt1.o is rv64i2p1_m2p0_a2p1_f2p2_d2p2_c2p0_zicsr2p0_zmmul1p0. Clearly, bwaves.lto.o is generated incorrectly. Could you please advise me on how to resolve this issue? Thank you very much!

The complete run script is：
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o block_solver.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 block_solver.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o flow_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 flow_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o flux_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 flux_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o jacobian_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 jacobian_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o shell_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 shell_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -fuse-ld=ld.lld -L{RISCV}/lib/ block_solver.o flow_lam.o flux_lam.o jacobian_lam.o shell_lam.o -o bwaves

[llvmbot]

@llvm/issue-subscribers-flang-ir

Author: bernadate (buggfg)

I am cross-compiling the **SPEC CPU2006** benchmarks using the **LLVM 19.1.0** compiler on an Intel machine. The C/C++ examples can be successfully built into executable files using runspec and run successfully on the** RISC-V** platform. However, for Fortran examples like **410.bwaves**, building with flang-new fails, and the error is:

ld.lld: error: bwaves.lto.o: cannot link object files with different floating-point ABI from ~/gnu-riscv64-20241106/sysroot/usr/lib/crt1.o.

PS. config info
**FOPTIMIZE **= -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math
**EXTRA_LDFLAGS **= -fuse-ld==ld.lld -L${gnu-riscv64}/lib/

Through debugging, I found that the ABI of bwaves.lto.o is rv64i2p1, while the ABI of the library file crt1.o is rv64i2p1_m2p0_a2p1_f2p2_d2p2_c2p0_zicsr2p0_zmmul1p0. Clearly, bwaves.lto.o is generated incorrectly. Could you please advise me on how to resolve this issue? Thank you very much!

The complete run script is：
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o block_solver.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 block_solver.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o flow_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 flow_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o flux_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 flux_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o jacobian_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 jacobian_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -c -o shell_lam.o -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -DSPEC_LP64 shell_lam.f
$ flang-new --target=riscv64-unknown-linux-gnu --gcc-toolchain={RISCV} --sysroot={RISCV}/sysroot -march=rv64gcv_zba_zbb_zbc_zbs -O3 -static -flto -ffast-math -DSPEC_CPU_LP64 -fuse-ld=ld.lld -L{RISCV}/lib/ block_solver.o flow_lam.o flux_lam.o jacobian_lam.o shell_lam.o -o bwaves

[jrtc27]

-march= isn't ABI, what matters is -mabi=, which has various defaults (based on ISA and OS). In the normal Unix rv64gc (plus possible extensions) case, that's lp64d. What does readelf -h say for the flags, as that's where the ABI is stored? Presumably Flang has failed to reimplement Clang's ABI defaults.

[jrtc27]

NB:

llvm-project/clang/lib/Driver/ToolChains/Arch/RISCV.cpp

Line 188 in 602bbf2

StringRef riscv::getRISCVABI(const ArgList &Args, const llvm::Triple &Triple) {

is Clang's implementation

[topperc]

Does it work without flto?

[ilovepi]

Before @topperc improved the target feature support, we'd hit this all the time in LTO builds with clang or rustc. The prevailing advice was to also pass the required -march=/-mabi= flags to the linker as well, via -Wl,-plugin-opt=. To unblock yourself you could try that, but I think @jrtc27's analysis is likely on point and fixing that will probably solve the issue for flang.

[buggfg]

@topperc Hello, in the SPEC CPU2006 benchmark , the Fortran programs cannot be successfully compiled in FLTO mode. However, some examples, such as 465.tonto, can run when FLTO is disabled. The above conclusion is based on personal observations.

[buggfg]

@jrtc27 @ilovepi Thank you for your attention！Yes, the Clang frontend has customized specific "target-abi" and "riscv-isa" information for the RISC-V architecture, and the RISC-V backend take these as the correct ELF attributes. However, the Flang frontend lacks this, which leads to the error. Currently, I have fixed this issue by applying a patch to the Flang frontend. I'm not sure if the Flang-LLVM community needs this small patch?

[rofirrim]

@buggfg I think we want to add first some plumbing in flang (driver) so it handles -mabi and passes -target-abi onto flang -fc1 like clang does with clang -cc1 (maybe you already did that).

Happy to help here.